Vial sort← Latest work

The August datasets

Versions 1 to 5, all public on Hugging Face in the LeRobot format. Any episode opens in the browser with all its cameras, the joint trajectory and the prompt on one timeline. The October set is on the Data page.

The demonstrations behind the evaluated policy

150 teleoperated demonstrations, about 58 minutes of recording: a person drives the leader arm while the cameras, the joint positions and the prompt are logged at 30 frames a second. Each demonstration is one move.

The set is sari-abdan/vial-sort-v5 on Hugging Face, and every episode can be watched in the browser.

A teleoperated demonstration from the v5 set. A person drives the leader arm; the follower copies it.

How the demonstrations were recorded

  • Three vial colours, red, cyan and green, three vials of each.
  • Two six-slot racks at a diagonal to the arm, the bin on the left.
  • Places are named from the side camera: the left rack is the one by the bin, and slots run 1 to 6 from left to right.
  • Vials are spaced so no two stand side by side, at most three to a rack; the other vials are never the colour being moved.
  • Slots 2 and 5 are never a source, so the evaluation can test two positions no demonstration started from.
  • The depth stream is recorded with every episode and not used by the policy.

Every demonstration has its own wording

150 wordings for 150 demonstrations, so the policy has to read the positions rather than memorise one sentence. The 8 for R1 → bin:

  • “Bring the tube at hole 1 of the rack on the right into the trash can.”
  • “Carry the test tube in position 1 of the right-side rack to the bin.”
  • “Move the sample tube from slot 1 of the right tube rack over to the trash.”
  • “Move the vial from position 1 of the right rack to the bin.”
  • “Relocate the test tube from spot 1 of the right-side rack to the rubbish bin.”
  • “Take the sample tube at the first slot of the right tube rack into the waste bin.”
  • “Take the vial in slot 1 of the right rack to the trash.”
  • “Transfer the tube from the first slot of the rack on the right over to the waste bin.”

Why the policy uses no depth

The vials are translucent plastic. The RealSense camera’s depth loses them; a model that estimates depth from the colour image alone keeps their shape. The policy is trained on the colour images.

One side-camera frame three ways: the colour image, the Depth Anything V2 estimate with every vial outlined, and the RealSense depth with the vials missing.
The same frame: colour image, Depth Anything V2, Intel RealSense.

Every dataset

Training datasets on the Hub

Each dot is one set of recordings; two names on a dot are the same recordings in two formats (joint and end-effector).

Show the numbers
DatasetUploadedEpisodesMinutesCameras
vial-sort-v1-static7 Jun 202610036cam_top, cam_wrist, cam_sidewatch
vial-sort-v1-ee8 Jun 202610036cam_top, cam_wrist, cam_sidewatch
vial-sort-v2-static23 Jun 202612040cam_top, cam_wrist, cam_sidewatch
vial-sort-v4-new29 Jun 20264012cam_top, cam_wrist, cam_sidewatch
vial-sort-v5-pilot10 Aug 202615058cam_top, cam_wrist, cam_side, cam_depthwatch
vial-sort-v522 Aug 202615058cam_top, cam_wrist, cam_side, cam_depthwatch
vial-sort-slots-10014 Oct 202610033top_down, left_top, left_bottom, right, wristwatch
vial-sort-contrast-5110 Nov 20265120top_down, left_top, left_bottom, right, wristwatch
vial-sort-slots-22010 Nov 202622079top_down, left_top, left_bottom, right, wristwatch
vial-sort-contrast-99-depth10 Nov 20269934top_down, left_top, left_bottom, right, wrist, rs_depthwatch